trl

Run `npx skills add huggingface/trl --all -g -y` in your terminal to install all tools in this suite globally.

TRL post-trains language models using methods like SFT, DPO, GRPO, and KTO. It helps you align models with human preferences or reward functions without writing training loops from scratch.

Load a prompt dataset, define a reward function, and pass both to GRPOTrainer with your model name. Enable vLLM with use_vllm=True to speed up generation during training.

Yes. TRL integrates with PEFT for LoRA and QLoRA, supports quantization, and works with DeepSpeed and FSDP for multi-GPU or multi-node scaling.

Yes. The built-in CLI lets you run commands like `trl sft` or `trl dpo` with your model and dataset names, no custom code required.